# Oliver AI raises undisclosed pre-seed to build a database for AI agents > Source: > Published: 2026-08-26 15:07:32+00:00 # Oliver AI raises undisclosed pre-seed to build a database for AI agents **Menlo Ventures and Unusual Ventures backed Praneet Sharma, Grace Johnson and Toby O'Brien as Oliver AI tests its performance case against internal benchmarks.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [PR Newswire](https://www.prnewswire.com/news-releases/menlo-ventures-and-unusual-ventures-back-oliver-ai-to-redefine-data-infrastructure-for-agentic-ai-302860322.html) ## Why it matters AI agents can turn routine analysis into a continuous stream of database queries. Oliver AI is betting that controlling that compute and access layer will become its own infrastructure market. [Praneet Sharma](https://www.linkedin.com/in/praneet-sharma-48262384?ref=runtimewire), Grace Johnson and [Toby O'Brien](https://www.scu.edu/alumni/connect/industry-based-communities/scu-software-industry-community/?ref=runtimewire) have raised pre-seed funding from [Menlo Ventures](https://menlovc.com/?ref=runtimewire) and [Unusual Ventures](https://www.unusual.vc/?ref=runtimewire) for [Oliver AI](https://oliverdb.ai/?ref=runtimewire), according to an [August 26 announcement](https://www.prnewswire.com/news-releases/menlo-ventures-and-unusual-ventures-back-oliver-ai-to-redefine-data-infrastructure-for-agentic-ai-302860322.html?ref=runtimewire). The founders are building an analytical database around a costly premise: software agents query data faster, more frequently and less predictably than the humans and dashboards that shaped the current generation of warehouses. Oliver AI did not disclose the round's size, valuation or lead investor. Menlo partner [Tim Tully](https://menlovc.com/team/tim-tully/?ref=runtimewire), formerly Splunk's chief technology officer, and Unusual Ventures co-founder [John Vrionis](https://www.unusual.vc/team/john-vrionis/?ref=runtimewire) both commented on the financing. They are financing an infrastructure bet well before Oliver AI has published customer, revenue or production deployment figures. Sharma has been working toward this problem through observability. Before Oliver AI, he co-founded Omlet, where his [technical writing](https://blog.omlet.co/archive?ref=runtimewire) covered OpenTelemetry data lakes, AI agents for site reliability engineering and Model Context Protocol integrations. Earlier roles included engineering and field work at Datadog, Clarifai and New Relic. That history explains Oliver AI's focus on the unglamorous part of agent deployment: the volume of queries, logs and tool calls generated after a model reaches production. O'Brien brings the commercial half of the founding team. A Santa Clara University accounting graduate, he previously led sales and software work at Datadog. ### The benchmark is the pitch [OliverDB](https://oliverdb.ai/?ref=runtimewire) combines a columnar analytical engine, an orchestration layer for groups of smaller models and policy controls for agents accessing databases or MCP servers. It accepts connections through SQL, REST and MCP. The founders are leading with speed and cost. Oliver AI's [benchmark notes](https://oliverdb.ai/benchmarks.html?ref=runtimewire) report speedups of up to 162x on CPUs and 1,000x on GPUs for evaluated analytical query shapes, along with compute reductions ranging from 50x to 300x depending on workload and configuration. Sharma says a multi-petabyte environment using thousands of servers could shrink to a few servers or one GPU. Those are internal results, not independently audited production measurements. Oliver AI compared its engine with ClickHouse 24.8 using query shapes from [ClickBench](https://github.com/ClickHouse/ClickBench?ref=runtimewire), an open benchmark covering clickstream analysis, web analytics, machine-generated data, logs and events. Oliver AI says hardware, concurrency, cache state, data shape and query mix can materially change the outcome. One example on Oliver AI's site compares an 8-millisecond OliverDB query with 3.8 seconds for a cloud warehouse, but the benchmark page says that figure represents a single analytical query rather than an end-to-end agent task. That distinction determines how customers should read the headline numbers. Agents can generate repeated exploratory queries while examining schemas, testing hypotheses and correcting earlier work. Faster individual queries could reduce both latency and infrastructure spending. The cost advantage will depend on how OliverDB behaves under sustained concurrency, mixed workloads and the security checks required inside a production environment. ### Giving agents narrower keys Oliver AI is selling governance alongside query performance. Its [Agent Flight Control](https://oliverdb.ai/governor.html?ref=runtimewire) product assigns a policy to each agent and rewrites SQL or MCP tool calls before execution. An enterprise can restrict rows, mask columns, deny operations and preserve an audit trail tied to a particular agent or user. The design targets a practical enterprise problem. An agent may need to read a customer table, inspect logs and open a ticket without receiving broad credentials to every underlying system. Oliver AI says its policies can travel with an agent's key across Postgres, Snowflake, ClickHouse and MCP tools. OliverDB can run inside a customer's virtual private cloud or as a managed service, allowing buyers with stricter data controls to keep information within their own network. Oliver AI also proposes using many smaller, specialized models in parallel. Its model swarm tests competing explanations and passes the findings to a conductor model, which selects the conclusion supported by the available data or returns no conclusion. The architecture is intended to reduce token spending and reserve larger frontier models for harder tasks. Oliver AI has not attached an external accuracy study to that claim, leaving enterprise evaluations to establish whether parallel model calls produce enough additional reliability to justify their operational complexity. ### The incumbent is already building for agents Oliver AI is entering a category that its benchmark target is helping define. ClickHouse has published an [Agentic Data Stack](https://clickhouse.com/blog/the-agentic-data-stack?ref=runtimewire) and launched [ClickHouse Agents](https://clickhouse.com/blog/clickhouse-agents-beta?ref=runtimewire) in public beta. Its stack connects agents to analytical data through MCP and adds tracing and evaluation tools around their interactions. That makes ClickHouse both Oliver AI's comparison point and an active competitor for the same workloads. Snowflake and Databricks are also adding AI interfaces and agent tooling around data already stored on their platforms. Oliver AI's opening is to sit alongside those systems, continuously ingest their data and give agents a faster, policy-scoped query layer without demanding an immediate replacement. Oliver AI advertises unlimited query compute for a flat monthly fee, although it does not publish pricing. Predictability could become a useful wedge if autonomous agents turn metered warehouse consumption into a budget problem. Buyers will still need production evidence showing that the savings survive ingestion, storage, networking, governance and peak concurrency. The pre-seed gives Sharma, Johnson and O'Brien room to turn an internally measured engine into customer proof. Their core observation is sound: a machine that can ask thousands of follow-up questions changes the economics of enterprise analytics. Oliver AI now has to show that its database can carry that workload outside the benchmark harness.